Diffusion Policy Eval

---
name: diffusion-policy-eval
description: Evaluates visuomotor policy learning in robotics by measuring how well a model generates sequential actions to complete manipulation tasks under both state and image observations. It probes the policy's ability to handle multimodal action distributions, long-horizon dependencies, and latency robustness across rigid and fluid object manipulation. Use when the user wants to benchmark on Robomimic, Push-T, Block Push, Franka Kitchen, or asks about evaluating this task. Reports success_rate.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2303.04137
  bibtex_key: [REDACTED]
  confidence: high
---

# diffusion-policy-eval

> Diffusion Policy: Visuomotor Policy Learning via Action Diffusion — Chi et al. (2023) (arXiv:2303.04137, 2023)

## What this evaluates

Evaluates visuomotor policy learning in robotics by measuring how well a model generates sequential actions to complete manipulation tasks under both state and image observations. It probes the policy's ability to handle multimodal action distributions, long-horizon dependencies, and latency robustness across rigid and fluid object manipulation.

## Datasets

- **Robomimic, Push-T, Block Push, Franka Kitchen** — total ?; splits: train (-1), test (-1)

## Metrics

- `success_rate` **(primary)** — range: [0, 1]
  - Fraction of rollout episodes where the task goal is successfully achieved. Reported as the average over the last 10 training checkpoints, 3 training seeds, and 50 environment initializations.
- `target_area_coverage` — range: [0, 1]
  - Percentage of the target area covered by the pushed block, used specifically for the Push-T task.
- `block_push_frequency` — range: [0, 1]
  - Frequency px of successfully pushing x blocks into targets in the Multimodal Block Pushing task.
- `kitchen_interaction_frequency` — range: [0, 1]
  - Frequency px of interacting with x or more objects in the Franka Kitchen task.

## Input / output format

**Input**: RGB images and/or proprioceptive state vectors (e.g., end-effector position, object poses) per timestep.

**Output**: A sequence of actions (position or velocity control) for a receding horizon, typically 8 steps ahead.

## Scoring recipe

```python
def compute_success_rate(predictions, gold, env_init=50, checkpoints=10):
    successes = 0
    total = 0
    for init in range(env_init):
        for ckpt in checkpoints:
            actions = predict_policy(ckpt, observations)
            success = run_episode(actions, gold)
            if success: successes += 1
            total += 1
    return successes / total
```

## Common pitfalls

- Using velocity control for Diffusion Policy instead of position control, which significantly degrades performance compared to the paper's optimal setup.
- Averaging over fewer than 50 environment initializations or not using the last 10 checkpoints, leading to inconsistent success rate estimates.
- Confusing short-horizon multimodality (multiple ways to achieve an immediate goal) with long-horizon multimodality (arbitrary sub-goal ordering), which require different evaluation metrics.

## Evidence (verbatim from paper)

> The metric for most tasks is success rate, except for the Push-T task, which uses target area coverage. In addition, we report the average of best-performing checkpoints for robomimic and Push-T tasks to be consistent with the evaluation methodology of their respective original papers. We report results from the average of the last 10 checkpoints (saved every 50 epochs) across 3 training seeds and 50 environment initializations

## Citation

```bibtex
@misc{chi2023diffusionpolicy,
  title={Diffusion Policy: Visuomotor Policy Learning via Action Diffusion},
  author={Chi et al. (2023)},
  year={2023},
  note={arXiv:2303.04137}
}
```

- arXiv: 2303.04137